Skip to content

[NPU] Avoid FP8 dispatch for INT8 MoE on A5 #40541

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
McZyWu wants to merge 6 commits into
base: main
Choose a base branch
from
+31 −4
Open

[NPU] Avoid FP8 dispatch for INT8 MoE on A5 #40541

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
9fbb553
[NPU] Avoid FP8 dispatch for INT8 MoE on A5
McZyWu Sep 21, 2026
0eed516
Merge branch 'main' into fix/a5-int8-deepep-dispatch
McZyWu Sep 24, 2026
bcff3cf
Merge branch 'main' into fix/a5-int8-deepep-dispatch
sglang-npu-bot Sep 28, 2026
7faa283
[NPU] Disable indexer top-k replay tests in CI
McZyWu Sep 29, 2026
a5a204c
Merge branch 'main' into fix/a5-int8-deepep-dispatch
sglang-npu-bot Sep 29, 2026
3feac03
Merge branch 'main' into fix/a5-int8-deepep-dispatch
sglang-npu-bot Sep 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions python/sglang/srt/layers/moe/token_dispatcher/deepep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,6 +490,15 @@ def set_deepep_dispatcher_dtype(self) -> None:
self.use_mxfp8 = config["use_mxfp8"]
self.use_nvfp4 = config["use_nvfp4"]

if _is_npu and self.deepep_output_dtype == DispatcherOutputDtype.INT8:
from sglang.srt.hardware_backend.npu.utils import is_npu_arch35

if is_npu_arch35():
# DeepEP interprets use_fp8=True as FP8 on A5, not INT8.
# Allow the normal-mode INT8 env fallback, or dispatch BF16
# and let the existing MoE kernel quantize it to INT8.
self.use_fp8 = False

# Handle environment variables
if _is_npu:
self._update_int8_quant_env()
Expand Down
13 changes: 11 additions & 2 deletions test/registered/npu/rl/indexer_replay/test_npu_return_indexer_topk_deepseek_v4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,8 +21,17 @@
)
from sglang.utils import wait_for_server

register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(
est_time=400,
suite="nightly-2-npu-a3",
nightly=True,
disabled="Disabled in CI; retained for manual execution.",
)
register_npu_ci(
est_time=400,
suite="base-b-test-2-npu-a3",
disabled="Disabled in CI; retained for manual execution.",
)


# Keep the first 6 layers of the BF16 checkpoint. The first 6 V4-Flash
Expand Down
13 changes: 11 additions & 2 deletions test/registered/npu/rl/indexer_replay/test_npu_return_indexer_topk_glm52.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,17 @@
popen_launch_server,
)

register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(
est_time=400,
suite="nightly-2-npu-a3",
nightly=True,
disabled="Disabled in CI; retained for manual execution.",
)
register_npu_ci(
est_time=400,
suite="base-b-test-2-npu-a3",
disabled="Disabled in CI; retained for manual execution.",
)


NUM_INDEXER_LAYERS = 6
Expand Down
Loading